Scholarlink

<p>The complex maritime environment poses significant challenges to small-target detection. Factors such as small target size, frequent occlusion, and unstable lighting conditions collectively limit detection accuracy. Firstly, we design WODown, a novel down-sampling module specifically for water-surface target detection. Its unique three-branch structure effectively mitigates small-target feature degradation, offering a new approach to enhancing the feature representation of small water-surface targets. Secondly, we propose a novel C2f-GC module. It integrates gated convolutional network blocks into C2f, significantly improving the model’s generalization ability. Finally, we introduce a dynamic detection head (Dyhead) to expand the model’s receptive field, thereby effectively reducing missed detections and false positives of low-resolution small targets. Comprehensive experiments conducted on the WSODD and Flow benchmark datasets have demonstrated the effectiveness of our method, achieving mean average precision (mAP) scores of 85.2% and 80.1%, respectively. These results represent improvements of 5.8% and 2.2% over the baseline models. The proposed solution offers an efficient framework for maritime target detection in complex scenarios involving unmanned surface vehicles.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CSD-YOLO: An Enhanced YOLOv8-Based Detector for Maritime Small Targets in Complex Environments

  • 期刊论文

  • 发表日期：

    2025年6月4日

查看全文
  • Yuanchen Sun,
  • Huanli Pang,
  • Changhong Jiang

摘要

The complex maritime environment poses significant challenges to small-target detection. Factors such as small target size, frequent occlusion, and unstable lighting conditions collectively limit detection accuracy. Firstly, we design WODown, a novel down-sampling module specifically for water-surface target detection. Its unique three-branch structure effectively mitigates small-target feature degradation, offering a new approach to enhancing the feature representation of small water-surface targets. Secondly, we propose a novel C2f-GC module. It integrates gated convolutional network blocks into C2f, significantly improving the model’s generalization ability. Finally, we introduce a dynamic detection head (Dyhead) to expand the model’s receptive field, thereby effectively reducing missed detections and false positives of low-resolution small targets. Comprehensive experiments conducted on the WSODD and Flow benchmark datasets have demonstrated the effectiveness of our method, achieving mean average precision (mAP) scores of 85.2% and 80.1%, respectively. These results represent improvements of 5.8% and 2.2% over the baseline models. The proposed solution offers an efficient framework for maritime target detection in complex scenarios involving unmanned surface vehicles.